How Is Macro News Transmitted to Exchange Rates?

March 21, 2006

Martin D. D. Evans1 Richard K. Lyons

Georgetown University and NBER U.C. Berkeley and NBER

Department of Economics Haas School of Business

Washington DC 20057 Berkeley, CA 94720-1900

Tel: (202) 338-2991 Tel: (510) 643-2027

evansm1@georgetown.edu lyons@haas.berkeley.edu

Abstract

Using data that avoid the stale quotes problem, we show that order flow contributes more to currency price

changes following public news than at other times. This is inconsistent with news effects being common

knowledge and therefore impounded in price directly. Roughly two-thirds of the total effect of macro news

on the DM/$ price is transmitted via order flow. With both the direct and indirect channels operating, macro

news accounts for 36 percent of total price variance, an order of magnitude more than earlier estimates.

Keywords: Information, Microstructure, Announcements, Heterogeneity.

1Respective affiliations are Georgetown University and NBER, and UC Berkeley and NBER. This paper was circulated underthe previous title, “Why Order Flow Explains Exchange Rates.” We thank the following for helpful comments: Brian Boyer,Anusha Chari, Frank Diebold, Jon Faust, Simon Gervais, Aditya Kaul, David Levine, Antonio Mello, Michael Melvin, MaureenO’Hara, Paul Pfleiderer, John Rogers, Andrew Rose, two anonymous referees, and seminar participants at the American FinanceAssociation meeting, the NBER Microstructure meeting, U.C. Berkeley, Alberta, Michigan, London Business School, the NewYork Fed, Aix-Marseille, Virginia, Wharton, Insead, LSE, the Central Bank of Norway, and the Central Bank of Italy. Bothauthors thank the National Science Foundation for financial assistance, including funding for a web clearinghouse for micro-basedFX research (see faculty.haas.berkeley.edu/lyons, georgetown.edu/faculty/evansm1).

1 Introduction

All textbook models of currency prices imply that public news determines prices directly: currency demand

shifts are common knowledge and any related transactions play no role in causing the change. In microeco-

nomic models of asset prices, transactions do affect prices causally (e.g., Glosten and Milgrom 1985, Kyle

1985). The causal role arises because transactions convey information that is not common knowledge. We

test whether transactions transmit macroeconomic news to currency prices, and how this channel compares

to the direct channel.

We examine the impact of macro news both intradaily and daily. We begin at the 5-minute frequency.

Estimates of our intraday model using interdealer order flows show that while order flow contributes signifi-

cantly to changing prices at all times, it contributes more to changing prices immediately after news arrival.2

This is inconsistent with the textbook view that macro news effects are common knowledge and therefore

impounded in prices without any order flow role. It suggests, instead, that macro news triggers trading that

reveals dispersed information, which in turn affects currency prices.

Our daily analysis provides further evidence that trading on news reveals incremental information. The

daily model distinguishes three sources of currency price variation. The first source mirrors traditional

models — macro news that is impounded immediately and directly. The second source is the indirect effect

of news on price via induced order flow. The third source is order flow that affects price but is unrelated

to public news (possibly induced by banks’ changing risk tolerances, firms’ changing hedging demands, or

individuals’ changing liquidity demands; see, e.g., Evans and Lyons 2002a). We find that all three sources of

DM/$ price variation are significant. Order flows vary considerably with macro news, with the result that

roughly two-thirds of the effect of macro news on currency prices is transmitted via order flow, the remainder

being the direct effect of news. This is consistent with the intraday finding that order flow is most important

for determining currency prices during periods immediately following news arrival. With both the direct and

indirect channels operating, we find that macro news accounts for 36 percent of total daily price variance.

This is an order of magnitude more explanatory power than in previous studies (addressed below).

Though the literature on news and currency prices is long standing, it has not used quantities (order

flow) to sort out the relationship. The literature has two branches: a first-moment branch that addresses

the direction of price changes and a second-moment branch that addresses price volatility. A common

finding of the first-moment branch is that directional price effects from scheduled macro announcements are

difficult to detect at the daily frequency — they are swamped by other factors. Intraday event studies, such as

Andersen et al. (2003), do find statistically significant effects, particularly for employment and money-supply

2Order flow is the cumulation over time of signed trades, where trades are signed according to whether the initiator is buyingor selling (the marketmaker posting the quote is the non-initiating side). Order flow’s role in determining currency prices isdocumented by Payne (2003), Rime (2000), Evans and Lyons (2002a,b), and Evans (2002), among many others. Flows fromindividual end-user segments in currency markets are addressed in Lyons (2001), Froot and Ramadorai (2005), and Evans andLyons (2005), among others. Order flow is similarly important for prices in bond markets, which share many informational andstructural features with currency markets (see, e.g., Green 2004, Fleming 2003, and Brandt and Kavajecz 2004).

1

announcements.3 The second-moment branch focusing on news effects on volatility is partly a response to

difficulty in finding news effects on return first moments.4 This work finds that announcements do indeed

produce the largest price changes.

The two papers most closely related to our own are Green (2004) and Love and Payne (2004). Green

focuses on the bond market and documents a significant increase in the informational role of trading following

economic announcements. Information asymmetry increases following the release of public information in a

way consistent with, for example, the skilled information processor models of Kim and Verrecchia (1994,1997);

see also Kandel and Pearson (1995). The Love and Payne (2004) paper, concurrent with our own, uses order

flow to sort our macro news effects on currency prices. They also find that a substantial share of public

news is incorporated into prices via trading. Though there are many differences across the analyses, one

key difference is exposure to the stale quotes problem: Love and Payne use data from limit-order trading,

whereas we use data on bilateral direct transactions. If there is any staleness in existing limit orders when

unambiguous news arrives, the first market orders that sweep out stale limit orders will be correlated with

price change, but do not truly cause it (Carlson and Lo 2004 has some discussion of this causality concern).

The data on bilateral direct transactions we use are free from this staleness.

Our paper departs from earlier work on news in two main ways: we exploit a much broader set of macro

events and we employ a different identification strategy. The set of announcements used in past work is

limited to scheduled announcements, which account for only about 10 percent of the news arrivals that

appear on trading-desk news screens (Reuters Money Market Headline News). Our wider set gives us a more

complete picture of the joint dynamics of currency (FX) prices and order flows, but requires that we account

for the absence of complete data on ex-ante announcement expectations. For identification, like Rigobon and

Sack (2004) we base our estimation strategy on state-dependent heteroskedasticity.5 Specifically, we identify

the relative importance of direct and indirect news effects by allowing news to affect the variances of order

flow and price differently. This approach does not require data on ex-ante expectations. Instead, it requires

the weaker assumption that one can identify changes in the variance of macro information shocks. To ensure

the robustness of our results, we model these variations in several different ways (in both the intraday and

daily analysis).

Our finding that macro news accounts for more than one third of price variance helps to resolve a big

puzzle in international finance — the news puzzle. The puzzle is that even the most comprehensive studies

of news effects on currency prices account for less than 5 percent of total price variation. A good example

3See also, for example, Cornell (1982), Engel and Frankel (1984), Hakkio and Pearce (1985), Ito and Roley (1987), Hardouvelis(1988), Klein (1991), and Ederington and Lee (1995). For bond markets, see Fleming and Remolona (1997) and Balduzzi, Elton,and Green (2001).

4 See, for example, Goodhart et al. (1993), DeGennaro and Shrieves (1997), Andersen and Bollerslev (1998), and Melvinand Yin (2000). For bond markets, see Fleming and Remolona (1999), Bollerslev, Cai, and Song (2000), and Huang, Cai, andWang (2002).

5 See the discussion in Rigobon and Sack (2004) comparing the merits of the event-study and heteroskedasticity approaches.Omitted variable bias in event-study analysis is simply a manifestation of a point made above, namely, that event effects areoften swamped by other factors affecting price.

2

at the daily frequency is Klein (1991). He regresses FX price changes on trade-balance news and finds that

news explains about 40 percent of price changes on those days. This is an impressive finding. However, since

trade balance news arrives monthly, roughly 95 percent of FX price variation is not included in the regression

(20 of 21 trading days per month). Thus, an R2 statistic of 0.4 implies that less than 3 percent of total price

variation is accounted for. Andersen et al. (2003) also report impressive R2 statistics within their event

windows (in this case, intraday windows). But as they note (p. 50), summing the amount of time in all of

their five-minute, post-event windows accounts for only 0.2 percent of their full sample period (e.g., roughly

one five-minute interval per day). Under the conservative assumption that news arrival causes variance to

increase by a factor of 10, their findings imply that news accounts for no more than 2 percent of the total

price variation.6 We estimate the contribution of macro news to be an order of magnitude higher for two

main reasons. First, our joint use of price and quantity data opens a new channel through which macro news

can effect FX prices. Second, as noted, we consider the effects of a much broader set of macro news events.

While scheduled announcements are certainly an important source of macro news, our estimates indicate

that they do not encompass all the sources of macro news available to market participants.

The remainder of the paper is in four sections. Section 2 describes our data and presents some descriptive

statistics. Section 3 presents the intraday analysis. Daily analysis is presented in Section 4. Section 5

concludes.

2 Data and Descriptive Statistics

Our order flow and price data are drawn from time-stamped, tick-by-tick transactions in the DM/$ spot

market over a four-month period, May 1 to August 31, 1996. The transactions are from the Reuters Dealing

2000-1 system that operates 24 hours a day, 7 days a week. Importantly, Dealing 2000-1 is a bilateral

interdealer system on which a dealer requests a quote from another dealer, and when received, generally has

only a few seconds to act before the quote is retracted. This type of data avoids the stale quote problem that

can cloud inferences about causality when news arrives since, unlike limit orders, these quotes are always

very short lived, and are generally not extended at moments of anticipated public news arrival. In 1996

at the time of our sample, Dealing 2000-1 was the most widely used electronic dealing system: according

to Reuters, over 90 percent of the world’s bilateral transactions between DM/$ marketmakers took place

through the system. Transactions between marketmakers accounted for about 75 percent of total trading in

major spot markets at the time. This 75 percent breaks into two transaction types–direct (bilateral) and

6Security-return volatility is not constant over time (French and Roll 1986). Our daily-frequency example from Klein(1991) could include two adjustments in this respect: currency price volatility over weekends is not zero, which lowers hisoverall explanatory power, but announcement days tend to have higher volatility than non-announcement days, which raiseshis explanatory power. Neither of these adjustments is large enough to alter the basic message. Andersen and Bollerslev (1998)report that Employment Report has the largest impact on the instantaneous variance, increasing it by a factor of 10. If allannouncements had the same effect, and the within-event-window R2 statistics were all one, news would still only account for 2percent of the total exchange rate variation. In fact, the R2 statistics in Andersen et al. (2003) are generally below 50 percent(Table 2), so the 2 percent figure is indeed an upper bound.

3

brokered (multilateral). Direct trading accounted for about 60 percent of trades between market-makers

and brokered trading accounted for about 40 percent. (For more detail on this Reuters Dealing System see

Lyons 2001 and Evans 2002; the latter includes details on data collection and statistical properties.) For

every trade executed on D2000-1, our data set includes a time-stamped record of the transaction price and

a bought/sold indicator. The bought/sold indicator allows us to sign trades for measuring order flow.

Our intraday analysis uses transaction prices, order flow and trade intensity measured over fixed intervals

of five-minutes. We denote the last DM price for the purchase and sale of dollars in interval i as paski and pbidirespectively. (With roughly 1 million transactions per day, the preceding transaction is only seconds before

the end of each 5-minute interval during regular trading hours.) Interdealer order flow, xi, is the difference

during interval i between the number of trades initiated by dealers buying dollars and the number initiated

by dealers selling dollars.7 Similarly, we measure trade intensity, ni, by the unsigned number of interdealer

transactions per minute during interval i. Daily versions of our series are denoted with subscript t for the

daily price pt, we use the last DM price for the purchase of dollars before 5 pm BST (British Summer Time).8

Daily order flow, xt, is the same as five-minute order flow xi save that it spans the time difference between

5 pm on days t− 1 and t. Trading intensity on day t, nt, is defined as the number of transactions over the

same daily interval.

The primary source of our news data is the Reuters Money Market Headline News screen (archived by

Olsen Associates). These screens are standard equipment on FX trading desks and are used for high frequency

monitoring by non-marketmaker participants as well. Reuters collects news reports from approximately 150

bureaus around the world. Each report must be approved by an economics editor at Reuters before it appears

as a news item on the Headline screens. The presence of this editorial process means that all the news items

in our data set were viewed as containing news-worthy economic information. At the same time, competition

between Reuters, Bloomberg and Dow Jones insures that editorial decisions minimize publication delay. We

impose a further layer of editorial screening by excluding from our data set news items of the following four

types: (i) reports of upcoming known holidays, (ii) reports that a scheduled data release would take place

(e.g., “Monthly employment report due out tomorrow”), (iii) duplicate reports (the same news is repeated

with a slight change in wording), and (iv) reports referring to the DM/$ price or market. The four filters

exclude less than 10 percent of news arrivals. The first three filters are intended to distill information that

is truly incremental.

A number of other factors give us confidence that our analysis is not significantly exposed to feedback

from the DM/$ market to macro news flow. The potential here is that increased volatility in the DM/$

price creates incentives for reporters to initiate news items to explain it, which are then posted to the

7 In direct trading between marketmakers, order sizes are standardized, so variation in size is much smaller than variation inthe size of individual trades between marketmakers and their end-user customers. Note too that using measures of order flowbased on numbers of transactions rather than size is common in work on equity markets, even when both measures are available(see, e.g., Hasbrouck 1991). Our data set does include total dollar volume over our sample, which allows us to calculate anaverage trade size, which we use below to interpret the estimated coefficients.

8Using prices from buyer-initiated transactions eliminates return reversals from prices bouncing randomly from bid to ask.

4

Headline screen. Our fourth filter helps to protect against this form of endogeneity insofar as the news

item makes reference to the DM/$ market. The well-defined editorial process described also helps protect

against spurious news creation. Perhaps most important, the Headline screen is used by traders in many

markets (money markets, bond markets, currency markets, and others), so the audience is much wider than

just the DM/$ market. We find the hypothesis of feedback to news flow patently strained when it comes to

our analysis at the five-minute frequency: the intermittency of arrivals shown in Table 1 makes this kind of

ultra-high-frequency feedback hard to imagine.

Table 1: Macro News Sample

Date Time News

5/1/1996 13:05:22 MARCH U.S. LEADING INDICATORS SHOW ECONOMY EASING5/1/1996 14:00:50 U.S. MARCH CONSTRUCTION SPENDING ROSE 3.1 PCT5/1/1996 14:10:14 MARCH U.S. CONSTRUCTION SPENDING REBOUNDS STRONGLY5/2/1996 6:05:18 GERMAN MARCH IMPORT PRICES CLIMB 0.3 PCT M/M5/2/1996 8:33:10 BUNDESBANK DOES NOT PLAN NEWS CONFERENCE TODAY5/2/1996 9:48:20 GERMAN CALL MONEY FALLS BACK TO 3.30/40 PCT5/2/1996 10:50:08 BUNDESBANK LEAVES INTEREST RATES UNCHANGED5/2/1996 10:51:56 GERMAN MARCH INDUSTRIAL OUTPUT DATA DUE 1130 GMT5/2/1996 12:30:40 U.S. JOBLESS CLAIMS FELL IN LATEST WEEK5/2/1996 12:31:00 U.S. Q1 1996 REAL GDP ROSE 2.8 PCT5/2/1996 14:00:30 U.S. MARCH FACTORY ORDERS ROSE 1.5 PCT5/2/1996 15:01:54 U.S. Q1 GDP SURGE SEEN JUST A BLIP IN MODEST TREND5/2/1996 15:10:32 GERMAN EMPLOYER TO OPPOSE ANTI-WAGE DUMPING LAW5/3/1996 9:56:32 GERMAN CALL MONEY EASES SLIGHTLY AHEAD OF WEEKEND5/3/1996 12:30:38 U.S. APRIL NON-FARM PAYROLLS ROSE 2,0005/3/1996 12:31:00 U.S. MARCH PERSONAL INCOME ROSE 0.5 PCT5/3/1996 12:38:44 U.S. JOBLESS RATE LOWER BUT LABOR MARKET LOOKS WEAK5/3/1996 13:42:16 MARCH US INCOME DATA SHOW MODEST GROWTH, INFLATION5/3/1996 14:00:16 U.S. MARCH HOUSING COMPLETIONS ROSE 5.1 PCT

Notes: The figure shows the macro news arrivals on the first 3 days of our sample (May1 to May 3, from the total sample from May 1 to August 31, 1996). Time is measured asBritish Summer Time (BST).

For further perspective on the news data, Table 1 lists the filtered news items over the first three days

of the sample. These items clearly extend well beyond scheduled releases of macroeconomic data.9 If we

were to focus only on the small subset of data releases that are scheduled, as we do when checking the

robustness of our results, there would be many trading days in our sample without a single news event.

The use of our broader set shows that in fact these were days on which macro news did arrive. Table 1

9A wider set of news types is also considered by Eddelbuttel and McCurdy (1998), though their focus is exchange ratevolatility, not the joint dynamics of news, signed order flow, and signed price movements. Berry and Howe (1994) use thenumber of public releases as a news flow measure to address equity market volume and volatility.

5

illustrates another feature of our data: although we filter out duplicate news items, we retain items that

interpret previous information. For example, we retain the third item listed in Table 1 because it interprets,

rather than simply restates, the information on construction spending contained in the second item. Does

such an interpretation represent news? Clearly the 3.1% increase in construction spending could have been

interpreted by some as representing a strong rebound, but it seems far-fetched to assume that everyone

subscribing to Reuters held this view and recognized the unanimity of opinion. When there is anything

short of a unanimous interpretation of a data release, a subsequent news item providing interpretation will

contain new information to at least some agents. Our prior is that data releases rarely (if ever) meet this

unanimity requirement and so we retain the interpretive items in our data set.10

Importantly, the estimation strategy we adopt in both our intraday and daily analysis does not require

that every news item is equally important. As we detail below, all we require is that the news data can be

used to identify variations in the flow of macro news hitting the FX market. For this purpose we construct

several different measures of macro news flow: one based on the arrival rates of just US news items, one based

on German items, and one based on the arrival of both US and German items. We also construct measures

from the small subset of releases that are scheduled. For this subset we can combine the Reuters data with

survey data on ex-ante expectations (provided by Money Market Service), allowing us to construct estimates

of macro news flow from estimates of unexpected components.11 We use these different measures of macro

news flow to check the robustness of our estimation results. In particular, since the arrival of scheduled news

is by definition immune to possible feedback from FX price volatility to the arrival of unscheduled news,

comparing results using all news verses scheduled news provides a test for the presence of feedback.

Table 2 presents descriptive statistics for the variables used in intraday and daily analyses. The upper rows

of panel A report sample statistics for the daily change in FX prices, ∆pt, and the level of interdealer order

flow xt. There is no detectable serial correlation in either variable at the daily frequency: The estimated first

order autocorrelation in the ∆pt and xt series are 0.015 and -0.035, and are both statistically insignificant.

The remaining rows in panel A report statistics on four of our measures of macro news flow. Aust and Agmt

respectively denote the number of US and German news items appearing on the Reuters Headline screen

between 5:01 pm BST on day t−1 and 5 pm BST on day t. At is the daily arrival rate of all news computed

as Aust +Agmt . Ast denotes the arrival rate for the subset of scheduled news, defined as the number of scheduled

releases between 5:01 pm BST on day t − 1 and 5 pm BST on day t concerning the following variables for

the US: Non Farm Payroll, Durable Goods, Trade Balance, and Unemployment Claims, and for Germany:

Current Account, Employment, GDP, Import Prices, Industrial Production, M3, Manufacturing Output and

Orders, Retail Sales, Trade Balance, and the Cost of Living. As the table shows, the median arrival rate

10Note, also, that we retained the eighth item listed in Figure 1 because in Germany (unlike the U.S.) the timing of thistype of release is not regular. Arrival of this type of information can certainly affect the timing of participants’ position taking.There was no record of the news item containing the German Industrial Production data in the archive created from the newsscreens by Olsen Associates — we do not know why.11Event studies in FX, such as Andersen et al. (2003), regress returns (price changes) on innovations for individual data

types (e.g. Unemployment Claims) computed from the MMS data. We cannot study the impact of scheduled news at thisdisaggregate level because our four-month sample contains too few scheduled releases for each data type.

6

Table 2: Sample Statistics

Min 5% 25% 50% 75% 95% Max Std. Skew. Kurt.A: Daily Data∆pt -20.7 -11.9 -3.8 0.3 3.4 6.9 12.4 5.9 -0.8 3.8xt -449.0 -308.0 -61.0 8.0 91.0 186.0 339.0 136.4 -0.6 4.5

Aust 0 0 1 2 5 7 9 2.2 0.7 2.6Agmt 0 2 6 8 12 18 22 5.1 0.5 3.0Aallt 0 2 9 11 15 21 27 6.0 0.2 2.6Ast 0 0 0 1 2 4 7 1.5 1.6 5.4

B: Intraday Data∆pi -7.9 -1.4 -0.3 0.0 0.3 1.3 5.0 0.8 -0.2 7.4xi -72.0 -9.0 -2.0 0.0 3.0 9.0 69.0 5.6 0.1 12.6ni 2.0 2.0 6.0 12.0 21.0 44.0 212.0 15.5 3.4 23.2

AutocorrelationsLag = 1 2 3 4 5 6 12 18 24∆pi -0.31 -0.00 -0.00 -0.00 -0.00 0.00 0.00 0.02 0.00

(<.01) (0.35) (0.76) (0.79) (0.68) (0.23) (0.69) (0.06) (0.64)

xi 0.23 0.10 0.09 0.08 0.06 0.06 0.03 0.02 0.00(<.01) (<.01) (<.01) (<.01) (<.01) (<.01) (<.01) (0.01) (0.65)

Notes: The sample is May 1 to August 31, 1996. ∆pt is 1000 times the change in the last DMpurchase price for dollars between 5:00 pm on day t and day t-1. xt is the total interdealer orderflow over the same time interval. Aust and A

gmt are respectively the number of macro news arrivals

observed on the Reuters Money Market Headline News screen relating to the US and Germanybetween 5:00 pm on day t and day t−1. Aallt denotes the total number of news items, Aust + Agmt ,and Ast is the total number of scheduled news items arriving over the same time interval. Scheduledannouncements include US Non Farm Payroll, Durable Goods, Trade Balance, and UnemploymentClaims and the German Current Account, Employment, GDP, Import Prices, Industrial Production,M3, Manufacturing Output and Orders, Retail Sales, Trade Balance, and the Cost of Living. Inpanel B, ∆pi and xi are the price changes (DM purchase price for dollars) and order flows during5-minute interval i, and ni is the total number of trades. Autocorrelations are computed by GMMas in Evans (2002) and the p-values reported in parenthesis are calculated from Wald tests of thenull hypothesis of a zero correlation (allowing for conditional heteroskedasticity).

for German news is four times the rate for US news. Notice also that the arrival rates for all news are

considerably higher than the rate for scheduled news.

Panel B of Table 2 presents descriptive statistics for prices, order flow, and trade intensity measured at

the 5-minute frequency. The sample statistics for ∆paski and ∆pbidi are almost identical, so we only report

those for ∆paski . One noteworthy feature of these statistics concerns the distribution of trade intensity, ni.

While the median trade intensity in our sample is 12 trades per minute, the distribution for ni indicates

that the pace of trading is occasionally much higher. Evans (2002) shows that some of the variations

7

in trade intensity can be related to the shift from predominantly Asian-based to US-based dealers as the

trading day progresses. However, on a particular day, variations in trade intensity can differ significantly

from this “seasonal” pattern. From the lower portion of panel B, we see a sharp difference from the daily

frequency statistics: both price changes and order flows are serially correlated at high (intraday) frequencies.

Transaction price changes (all are ask prices) display significant negative autocorrelation, but only at lag

one, while order flow appears serially correlated at up to 18 lags.

We track the arrival of news at the 5-minute frequency with dummy variables. The dummy variable Ai

takes the value of one if either a US or German news item appears on the Reuters screen during interval

i. At least one news arrival occurs in 515 out of the 11,473 five-minute windows in our four-month sample.

We use this dummy-variable approach in the five-minute data because there are few instances of more than

one news arrival during a single five-minute observation window (in 29/515 there were two arrivals and in

4/515 there were three, numbers that proved insufficient to get mileage from a multi-valued dummy). We

also make use of analogous dummies for all German news, all US news, and all scheduled US news; denoted

respectively by Agmi , Ausi and Asi .

3 Intraday Analysis

Our intraday analysis is based on a model for the joint dynamics of FX prices and order flows estimated at

the 5-minute frequency. Information is impounded into FX prices via two channels. The first is the direct

channel through which the arrival of new common-knowledge information leads marketmakers to change

the FX prices they quote. The transmission of information into FX prices via this channel is direct and

instantaneous. The second channel, the indirect channel, operates via order flow. In this case the arrival of

information is first manifest in the trading decisions of individuals because the information is dispersed. Once

marketmakers observe the ensuing order flow, they adjust their FX quotes to reflect the new information

embedded in the pattern of trading. Thus, order flow is the medium by which dispersed information becomes

embedded into FX prices.

Our intraday analysis will focus on the relative importance of the direct and indirect information channels

in the period immediately following the arrival of news. The motivation for this focus is straightforward:

If macro news primarily comprises new common-knowledge information, as is traditionally assumed, we

should find evidence that the direct channel accounts for most of the FX price variation over intervals that

include the news arrival. Conversely, if the arrival of macro news triggers revelation of dispersed information,

possibly reflecting diverse views about price implications, we should find that the indirect channel dominates.

We will quantify the relative importance of the direct and indirect channels from a decomposition of the

variance in FX price changes.

8

3.1 The Model

Our intraday model extends the empirical model in Evans (2002) to account for the effects of news arrivals.

At the heart of the model are the following equations:

∆pi = B(L)ξi + εi, (1)

yi = Cy(L)ξi, (2)

where ∆pi is the change in the spot price of FX between the end of periods i−1 and i, and yi is the order flowinitiated by end-users during period i. (The relationship between this end-user flow yi and inter-dealer flow

xi is addressed below.) Equation (1) shows how prices respond to two types of news: common knowledge

news shocks εi, and dispersed information shocks, ξi. We assume that these shocks are mutually orthogonal

and serially uncorrelated (addressed below). The εi shocks represent unambiguous price-relevant news that

is simultaneously observed by everyone and so are impounded fully and instantaneously into the price of

FX. Dispersed information shocks represent, in aggregate, the bits of information contained in the trades of

individual agents. This information is first manifested in the order flow, yi, and then subsequently impounded

in price. End-user order flow is the difference between the purchase and sales of dollars initiated by end-users

at dealer FX quotes. The dynamic responses of prices and order flow to these dispersed information shocks

are determined by the lag polynomials B(L) and Cy(L).

Three features of our specification deserve note. First, equation (1) describes the dynamics of transactions

prices, pi, defined as the market-wide average price at which actual transactions take place at time i.We will

describe the link between this pi and actual transactions below. Second, the assumed orthogonality between

the common-knowledge and dispersed information news shock implies that common-knowledge news has no

effect on order flow. This assumption has a long history in empirical finance, dating back at least to the

work of Hasbrouck (1991), and serving as the basis for much important work by various authors since then

(see, e.g., Madhavan, Richardson, and Roomans 1997 and the survey in Madhavan 2000). Intuitively, any

revision in price due to common-knowledge news should establish a new market-clearing price that does not

systematically favor subsequent imbalances of sell orders over buy orders, or vice versa. For example, there

should not be a correlation between bad public news for the DM and subsequent net DM sell orders, so

long as the update of the market price is unbiased. (Notice that this has nothing to do with the behavior of

unsigned trading volume; our model does not restrict how common-knowledge news affects volume through,

say, portfolio rebalancing.) The third feature concerns the dynamics of end-user order flow yi: We assume

that end-users’ demand for foreign currency is imperfectly elastic, so any imbalance in order flow (i.e., yi 6= 0)requires price adjustment to achieve market clearing. Consequently, all order flow is, at least temporarily,

price relevant.12 Under rational expectations, this information is summarized in current and past dispersed

12Our elasticity assumption does not imply that shocks to order flow necessarily have permanent price effects. It is possiblethat some shocks to order flow only affect prices while the associated inventory imbalance is being spread among dealers (see

9

information shocks, but remains unrelated to common-knowledge news shocks, as shown in equation (2).

Equations (1) and (2) allow us to identify three channels through which the arrival of macro news may

affect the dynamics of price and order flows. First, when the macro announcement contains a common-

knowledge component, it will affect prices instantaneously via the εi shock. This direct channel will be

operable when everyone agrees on the price-implications of the announcement. Second, when a macro

announcement is viewed by different agents as having different price implications, its effects on prices and

order flow will manifest via the ξi shocks: Although everyone observes the same announcement, different

views about the mapping from macro data to FX prices represent dispersed information that is relevant

for equilibrium prices. Third, the arrival of a macro announcement can affect the process through which

dispersed news is impounded into prices, by which we mean the lag polynomials. We allow for this by

allowing B(L) and Cy(L) to vary with the arrival of new announcements.

3.1.1 Empirical Specification

Estimation of our intraday model is complicated by two factors: First, our data are on market-wide order

flow between dealers, xi, rather than the end-user order flows yi. We must be careful to distinguish these

different order flows if we are to account for the temporal impact of dispersed information. Second, our

model needs to accommodate forms of state-dependency beyond the arrival of macro news. We shall deal

with these complications in turn.

Prices in the data set come in two forms. If a dealer initiating a transaction buys dollars, the transaction

price equals the ask quote in DMs per dollar offered by the other marketmaker. We refer to this as the

DM purchase price for dollars, pask. If the dealer initiating a transaction sells dollars, the transaction price

will equal the bid quote given by the other dealer. We refer to this as the DM sale price for dollars, pbid.

Evans (2002) finds evidence that lack of transparency in direct dealer trading allows for an equilibrium price

distribution, as opposed to a strict law of one price. To formalize this idea, our intraday model assumes

that equilibrium in the market at a point in time is described by a distribution of purchase prices and a

distribution of sales prices.

Let paski and pbidi denote observed prices drawn randomly from the respective distributions of purchase

and sales prices at time i. These observed prices are related to the average transaction price, pi, defined in

(1), by:

poi = pi + ηoi , (3)

for o = ask,bid. ηaski and ηbidi are idiosyncratic shocks that identify the degree to which observed prices

differ from the market-wide average. Their size depends on the identity of the dealers whose prices we

observe. We assume that observed prices are drawn randomly and independently from the cross-sectional

Cao, Evans and Lyons 2006). In this special case, some of the individual coefficients in B(L) will differ from zero, but theirsum will equal zero.

10

distributions of purchase and sale prices every period, so that ηaski and ηbidi are serially uncorrelated and

independently distributed.

The second complication arises from the distinction between the interdealer and end-user order flows.

The order flow measure in our data set is derived from trades initiated between dealers. These trades

are temporally downstream from the trades initiated by end-users against dealer quotes. As a result, it is

possible for a dispersed information shock ξi to affect prices and end-user order flows before it shows up in

interdealer order flow: Dealers may adjust their price in the face of an end-user order induced by ξi before

initiating trades in the interdealer market for risk sharing or speculative motives. Thus, price changes may

appear temporarily prior to changes in interdealer order flow even though they represent a response to earlier

end-user order flow. We allow for this possibility by assuming that the interdealer order flow we measure is

a distributed lag of end-user order flow:

xi = Cx(L)yi−m, (4)

where, again, Cx(L) is a polynomial in the lag operator. In this specification, it takes at least m periods

before imbalances in end-user orders for FX show up in interdealer order flow (where m may be zero).

Combining (4) with (1) and (2), we can now represent the dynamics of prices and interdealer order flow

by:

∆pi = D(L)xi + εi, (5)

xi = C(L)ξi−m (6)

where D(L) = B(L)L−mC(L)−1 and C(L) = Cx(L)Cy(L). Although the polynomial D(L) may take many

forms depending on the dynamic responses of price and interdealer order flow to dispersed information

shocks, in general it will include both negative and positive powers of L (corresponding to leads and lags

of xi) when m > 0. Our model estimates are based on a sixth-order specification for D(L) (shown below)

that links ∆pi to interdealer order flows from xi+4 to xi−1. This specification is supported by a series of

diagnostic tests reported in Evans (2002). It implies that a dispersed information shock may impact end-user

orders and prices up to 20 minutes before it affects interdealer order flow (i.e., m = 4). Similarly, we specify

the form of C(L) so that the time series properties implied by (6) match those in the data. As in Evans

(2002), we find that interdealer order flow is well characterized by an AR(10) process, so we specify C(L) as

(1−P10

j=1 cjLj)−1.

Finally, we incorporate the effects of macro news. We treat the arrival of news as changing the state of

the market. Following Evans (2002), we also allow the dynamics of prices and order flow to vary with trading

intensity. Including trading intensity as a state variable is important for accommodating the pronounced

time dependence in volatility documented by Andersen and Bollerslev (1998). Let Si denote the state of the

market in period i. We assume that Si depends on trading intensity in period i, ni, and the arrival of news

during the past three periods, Ai, Ai−1 and Ai−2. (Recall that the dummy variable Ai equals one if a macro

11

news arrives during period i.) We incorporate state-dependency into the price and order flow dynamics via

the polynomial D(L), and the error variances. Specifically, D(L) is replaced by D(L, S), a state-dependent

sixth order polynomial:

D(L, S) = d1(n, A)L−4 + d2(n, A)L

−3 + ....+ d5(n, A) + d6(n, A)L. (7)

where Ai ≡ max Ai, Ai−1, Ai−2 with state-dependent coefficients dj(., .). Thus, d6(n, 1) is the coefficienton lagged order flow xi when trade intensity equals n and news arrived in the past 15 minutes. We also

allow for state-dependence in the error variances, V ar(εi|Si) = Ωε(ni, Ai), V ar(ξi|Si) = Ωξ(ni, Ai), and

V ar(ηaski |Si) = V ar(ηbidi |Si) = Ωη(ni, Ai). State-dependence in the coefficients and variances is modeled as:

dj(n, A) = dj(A)e(−n/100) + dj(A)[1− e(−n/100)], (8)

Ωj(n,A) = ωj(A)e(−n/100) + ωj(A)[1− e(−n/100)], (9)

where dj(0), dj(0), ωj(0), and ωj(0) are the parameters to be estimated for observations without a news

arrival, and dj(1), dj(1), ωj(1), and ωj(1) when there is a news arrival. These functional forms make dj(.)

and Ωj(.) smooth monotonic functions of the transaction rate and are similar to the transition functions

used in nonlinear time series models (Potter 1999). They bound the coefficients between dj(A) and dj(A),

and the variances between ωj(A) and ωj(A) as the transaction rate varies between 0 and ∞.

Several aspects of our specification for state-dependency deserve comment. First, while specialized with

respect to variations in trading intensity, the functional forms in (7) - (9) do not appear unduly restrictive

when we subject our model to specification tests below. Second, there is no evidence that variations in trading

intensity or the arrival of news affect the dynamics of order flow via C(L). Thus, we do not incorporate state-

dependency in this polynomial to avoid an unnecessary proliferation in parameters. Third, our specification

places minimal restrictions on how the arrival of news affects the error variances and the link between order

flow and price dynamics. Importantly, we do not restrict how the coefficients inD(L, S) or the error variances

change following the arrival of news. Consequently, our specification does not impose a prior about how the

arrival of macro news affects the relative importance of the direct and indirect information transmission

channels. Finally, our specification makes no distinction between the arrival of US news, German news,

scheduled news or unscheduled news; Ai equals one when any news arrives during period i. We recognize

that this assumption may be too restrictive. For example, it is possible that the information transmission

process following the arrival of scheduled US news differs from that following other news items. Below we

investigate the adequacy of this assumption with a series of specification tests.

12

3.1.2 Estimation

The model is estimated using the Generalized Method of Moments technique developed in Evans (2002).

The moment conditions used to estimate the parameters of the order flow process are

0 = E [ξi ⊗ zxi ] , (10a)

0 = E£©ξ2i − Ωξ(Si)

ª⊗ zxi

¤, (10b)

where ξi = xi+4 −P10

j=1 cjxi+4−j and Ωξ(Si) is the conditional variance of ξi specified in (9). (Hereafter,

we use Si rather than ni and Ai as the argument of the error variances, Ω(.).) If the order flow process

is correctly specified, a dispersed information shock ξ in period i should be uncorrelated with interdealer

order flow x in periods i + 3 and earlier. Similarly, the difference between ξ2i and the conditional variance

should be uncorrelated with current or past trade intensity and order flows. We employ xi+3, ....xi−6 andfour lagged values of ξi as elements of the instrument vector z

xi in (10a). In (10b) the instrument vector

contains a constant, e(−ni/100) and Ai.With this choice of instruments, equations (10a) and (10b) represent

17 moment restrictions on 14 parameters (cj10j=1 ,ωξ(0),ωξ(1), ωξ(0) and ωξ(1)).

Parameters of the price process are computed from moments using the bivariate process for purchase and

sales prices, ∆paski and ∆pbidi . Combining (3) with (5) and our specification for D(L, S) gives:

∆poi =P6

j=1 dj(Ai)xi+5−j +P6

j=1[dj(Ai)− dj(Ai)]e−ni/100xi+5−j + uoi

where uoi ≡ εi + ηoi − ηoi−1 for o= ask,bid. This equation describes the state-dependent relation betweenactual transactions prices and interdealer order flow implied by our model. Notice that the composite error

term uoi follows an MA(1) process and that Cov (uaski , ubidi ) = Ωε (ni, Ai). We account for this error structure

in the moment conditions used to estimate the parameters of the price process:

0 = E [uoi ⊗ zpi ] (11a)

0 = E£©(uoi )

2 − Ωε(Si)− Ωη(Si)− Ωη(Si−1)ª⊗ zpi

¤(11b)

0 = E [uoiuøi − Ωε(Si)⊗ zpi ] (11c)

0 = E£©uoiu

oi−1 +Ωη(Si−1)

ª⊗ zpi

¤(11d)

0 = E£uoiu

øi−1 ⊗ zpi

¤(11e)

0 = E£uoiu

oi−2 ⊗ zpi

¤(11f)

0 = E£uoiu

øi−2 ⊗ zpi

¤(11g)

for o,ø = ask,bid and ø 6=o. The moment restriction in (11a) exploits the assumed orthogonality betweenthe instruments, zpt , and both the common knowledge news and idiosyncratic shocks. The other restrictions in

(11) are derived from the moving average structure of the composite error. In particular, (11b) and (11c) focus

13

on the variance of uaski , ubidi , while (11d) - (11g) focus on the the autocovariance. For example, in (11f) and(11g) we exploit the fact that under an MA(1) process, all the autocorrelations in the composite errors at lag

2 are zero. We use xi+j , e(−ni/100)xi+j , Aie(−ni/100)xi+j4j=−1 as instruments in (11a),

©1, e(−ni/100), Ai

ªin (11b) - (11d), and a constant in (11e) - (11f). This instrument choice gives us 66 moment restrictions on

the 32 parameters of the prices process (©dj(0), dj(1), dj(0), dj(1)

ª6j=1

and©ωj(0), ωj(1), ωj(0), ωj(1)

ªj=ε,η

).

Thus, the complete model contains 83 moment restrictions on 46 parameters.

In standard time series applications, GMM estimates of the parameter vector θ are found by minimizing

a quadratic form constructed from the sample analogues of the moment conditions implied by the model.

In this application, estimation is complicated by the fact that the gap between successive purchases and/or

sales occasionally span many minutes. In these cases there is no record of an FX purchase and/or sale in

the observation interval. For the purpose of computing our estimates, we designate the price, and order flow

observations from these periods as “missing” and construct sample moments without these observations.

Specifically, let E[mi,j(θ)] = 0 denote condition j among the moment conditions shown in (10) and (11) and

let Λ = i1, .i2...iT be the set of observations for which none of the elements in mi,j(.) for all j is “missing”.

We compute the sample analogue to condition j as mj(θ) = T−1PΛmi,j(.). The GMM estimates of θ are

then found by minimizing:

Q(θ) = m(θ)0W−1m(θ), (12)

where m(θ) = [m1(θ), m2(θ), ....]0. We follow the standard practice of first setting the weighting matrix W

equal to the identity to obtain consistent estimates of θ. These estimates, θ, then are used to compute a

consistent estimate of the optimal weighting matrix, W . We construct W using the Newey and West (1987)

estimator for the covariance of mi,j(θ) incorporating a correction for MA(1) serial correlation. This estimate

of the covariance matrix allows for the fact, documented below, that the model fails to completely account

for the heteroskedasticity in prices and order flow. The GMM estimates, θ, are found by minimizing (12)

with W = W . The asymptotic covariance matrix of the resulting estimates is V = [GW−1G0]−1 where

G = ∂m(θ)/∂θ0.

We examine the performance of our estimated model with a series of diagnostic tests. In particular, we

use a chi-squared test to examine the validity of an auxiliary set of moment conditions implied by our model

but not used in estimation. Let mii(θ) denote a vector of Kii sample moments, comprising the Ki moments

used to find the GMM estimates, and Kii−Ki auxiliary moment conditions implied by the model. Following

Hayashi (2000), we construct the test statistic by first finding the GMM estimates of θ, denoted θii, from

the set of Kii moments. These estimates are found with the two-step procedure described above using the

Newey and West estimator from the first step to construct the weighting matrix, Wii. Next, we construct the

submatrix of Wii corresponding to the original Ki moments, Wi. We then find an alternative set of GMM

estimates, θi, by minimizing (12) with W = Wi. Finally, we form the test statistic

C ≡ Tmii(θii)0W−1ii mii(θii)− Tm(θi)

0W−1i m(θi). (13)

14

where T denotes the number of “non-missing” elements used to construct mii(θ). Under the null hypothesis

that the auxiliary moment conditions are satisfied, the C statistic has an asymptotic chi-squared distribution

with Kii −Ki degrees of freedom. We use this test below to examine the adequacy of our specification for

the state-dependent coefficients and error variances.

3.1.3 Model Estimates

Table 3 presents GMM estimates of the intraday model. In specifications where all the variance parameters

where left unrestricted, the estimates of ωε(A), ωξ(A), and ωη(A) were very close to zero (i.e. < 0.0001), so

the table reports estimates where these parameters are restricted to zero. With these restrictions imposed,

there are 40 parameters to be estimated from a total of 64 moment restrictions, so our estimates are derived

from a model with 24 over-identifying restrictions. The Hansen (1982) J—statistic computed from our GMM

estimates is 55.722 which implies a p-value of 0.092 for the null of a correctly specified model.

Panel A of Table 3 reports the parameters for the state-dependent order flow polynomial, D(L, S). A

comparison of the estimates in rows (i) and (ii) and rows (iii) and (iv) shows that trade intensity has differing

effects on the price-impact of order flow depending on the arrival of news. This is most easily seen in the

right hand column where we report the sum of the coefficients in different market states. These estimates

have two noteworthy features. First, the long run impact of order flow on prices is much larger when trading

intensity is high (P

j dj(.) <P

j dj(.)). Second, controlling for trading intensity, the arrival of news slightly

reduces the long—run impact of order flow (P

j dj(A = 1) <P

j dj(A = 0), except at the very lowest trade

intensities). Further evidence on the importance of state-dependency is provided by the four test statistics

shown at the bottom of the panel. Here we report the results of Wald tests for the following coefficient

restrictions: (i) dj(0) = dj(0), (ii) dj(1) = dj(1), (iii) dj(1) = dj(0), and (iv) dj(1) = dj(0) for j = 1, ..., 6.As the table shows, there is strong statistical evidence against all of these restrictions. These findings are

consistent with the non-parametric evidence on state-dependence in hourly price change data reported in

Evans and Lyons (2002b). They also illustrate how intraday data brings greater resolution to the study of

price and order flow dynamics than is possible with models estimated at lower frequencies.

Parameter estimates from the order flow equation are reported in Panel B. Many of the coefficients are

highly statistically significant, indicating that there is indeed a good deal of serial correlation in intraday

order flow. The table also reports the estimate of (1 −P

j cj)−1 which measures the cumulative long-run

effect of dispersed information on order flow. The estimate of 1.69 indicates that the cumulative effect of a

dispersed information shock is approximately 70 percent greater than its initial impact.

Panel C of Table 3 reports the estimated parameters of the state-dependent error variances. The estimated

values for ωη(A) imply that the standard deviation of the idiosyncratic shocks slowly falls from approximately

8 to 3 per cent as n varies from 2 to 200. Thus, the cross-sectional dispersion of transactions prices falls as

trade intensity increases, as in Evans (2002), but we find no evidence that dispersion depends on the arrival

of news. The estimates of ωε(A) indicate how the volatility of common-knowledge shocks varies with trade

15

Table 3: GMM Estimates of the Intraday Model

A: Price Equation: ∆pi =P6

j=1

©dj(Ai)e

−ni/100 + dj(Ai)(1− e−ni/100)ªxi+5−j + εi

d1(.) d2(.) d3(.) d4(.) d5(.) d6(.)P

j dj(.)

A = 0 0.029 0.025 0.028 -0.047 -0.113 -0.034 -0.113(0.024) (0.057) (0.233) (0.052) (0.025) (0.033) (0.030)

A = 1 -0.022 0.074 0.054 -0.131 0.002 -0.066 -0.089(0.045) (0.046) (0.042) (0.046) (0.044) (0.043) (0.070)

d1(.) d2(.) d3(.) d4(.) d5(.) d6(.)P

j dj(.)

A = 0 0.127 0.275 0.543 0.629 -0.220 -0.062 1.293(0.106) (0.210) (0.716) (0.186) (0.078) (0.101) (0.106)

A = 1 0.278 -0.018 0.256 0.858 -0.449 0.091 1.015(0.153) (0.139) (0.131) (0.133) (0.107) (0.114) (0.209)

Wald Tests

dj(0)=dj(0) dj(1)=dj(1) dj(1)=dj(0) dj(1)=dj(0)216.083 19.096 20.896 11.953(<0.001) (0.004) (0.002) (0.063)

B: Order Flow Equation: xi =P10

j=1 cjxi−j + ξi−4

c1 c2 c3 c4 c5 c60.21 0.036 0.048 0.033 0.019 0.025(0.014) (0.013) (0.012) (0.012) (0.011) (0.011)

c7 c8 c9 c10 (1−P10

j=1 cj)−1

0.015 0.017 -0.016 0.020 1.688(0.010) (0.012) (0.010) (0.008) (0.070)

C: Variance Parameters: Ωj (n,A) =ωj(A)e−ni/100 + ωj(A)¡1− e−n/100

¢Shock Types

Idiosyncratic Common Knowledge Dispersed Informationωη(.) ωη(.) ωε(.) ωε(.) ωξ(.) ωξ(.)

A = 0 0.002 0.000 0.000 0.010 0.000 0.032(<0.001) (<0.001) (0.002)

A = 1 0.002 0.000 0.000 0.006 0.000 0.032(<0.001) - (0.002) (0.034)

Notes: The table reports GMM estimates with asymptotic standard errors in parentheses corrected forconditional heteroskedasticity and an MA(1) error term. News arrival is denoted by Ai and Ai, withAi = max Ai, Ai−1, Ai−2 where Ai = 1 if there was a news arrival during the previous 5-minutes.Coefficients and standard errors in panels A and C are multiplied by 100. P-values are reported in parenthesesbelow the Wald statistics in panel A. For the variance parameters, P-values are not reported in cases whereunrestricted parameter estimates were <0.0001 because these parameters were restricted to zero.

16

intensity and the arrival of news. The estimated standard deviation of common-knowledge shocks rises from

approximately 1 to 9 percent as n varies between 2 and 200 when news is absent, and from 1 to 7 percent

when news arrives. The estimated standard deviation of dispersed information shocks also increases with

trade intensity: from 0.03 to 0.17 as n varies between 2 and 200, whether or not news arrives.

Two implications of these estimates deserve emphasis. First, under normal trading conditions, much of

the observed volatility in high frequency transactions prices is attributable to the dispersion of prices that

characterizes market activity at a point in time.13 Failure to account for this feature of the data would leave

our analysis of how news arrivals affect prices and order flow flawed. Second, our estimates only show how

the arrival of news affects price and order flow dynamics for a given level of trade intensity. If the arrival of

news changes trade intensity, as indeed it does, the total impact of news on prices and order flow will reflect

both the direct effect of news and the indirect effects associated with the induced change in trade intensity.

We examine the combined effects of news in Table 5 below.

Table 4: Diagnostics for Intraday Model

(a) (b) (c) (d) (e)C(L) D(L,S) Ωε (S) Ωξ (S) Ωη (S)

Instrument: zi(i) Trade Intensity ni 2.624 0.371 0.737 0.200 0.134

(0.989) (0.999) (0.391) (0.655) (0.714)

(ii) US News Ausi 1.731 19.083 0.950 1.731 0.019(0.188) (0.087) (0.330) (0.188) (0.891)

(iii) Scheduled News Asi 17.905 13.084 3.904 2.307 1.660(0.084) (0.363) (0.068) (0.129) (0.198)

(iv) Residual ARCH 17.543 30.123 8.190(0.001) (<0.001) (0.042)

Notes: The table reports C-tests for a set of auxiliary moment conditions implied bythe model. In column (a) the restrictions take the form E[ξixi+4−jzi] = 0 for j =1, 2, .., 10. The restrictions in (b) are E[uoi zixi+5−j ] = 0 for j = 1, 2, ..6 whereuoi ≡ εi+η

oi−ηoi−1 for o = ask,bid. In columns (c) - (e) the restrictions are E[κizi] = 0

where κi ≡ uoiuøi − Ωε(ni, Ai) in (c), κi ≡ ξ2i − Ωξ(ni, Ai) in (d), and κi ≡ uoiu

oi−1 +

Ωη(ni−1, Ai−1) in (e). The instruments zi are ni, Ausi Asi and κi−j for j = 1, 2, 3 inrows (i) - (iv) respectively. P-values are reported in parentheses.

Our specification for the intraday model imposes many more moment conditions than were used in GMM

estimation. Table 4 provides diagnostics in the form of C-tests on a selection of these additional moment

conditions. The tests in column (a) look for state-dependency in the order flow polynomial C(L). For this

13Evans (2002) describes how the low degree of transparency of the FX institutional structure allows for the existence of aprice distribution without introducing arbitrage opportunities.

17

purpose we compute C-statistics for restrictions of the form E[ξixi+4−jzi] = 0 for j = 1, 2, ..., 10 ; wherezi equals ni, Ausi and Asi in rows (i) (ii), and (iii) respectively. These moment conditions will not hold

if, contrary to the assumption of our model, the serial correlation in order flow varies with either trade

intensity, the arrival of US news, or the arrival of scheduled news. The tests reported in column (b) look

for misspecification in the estimated form of the D(L, S) polynomial. In this case the restrictions being

tested take the form E[uoi zixi+5−j ] = 0 for j = 1, 2, ..., 6 where uoi ≡ εi + ηoi − ηoi−1 for o = ask,bid.These tests look for evidence of state-dependency in D(L, S) beyond that implied by functional form in

(7) and (8). Similarly, the C-tests in columns (c)-(e) look for evidence of misspecification in the error

variances. The moments being tested here take the form of E[κizi] = 0 where κi is the unexpected squared

realization of the shock in period i [i.e., κi ≡ uoiuøi − Ωε(ni, Ai) in column (c), κi ≡ ξ2i − Ωξ(ni, Ai) in (d),

and κi ≡ uoiuoi−1 +Ωη(ni−1, Ai−1) in (e)]. Row (iv) reports C-tests for 3rd order residual ARCH by testing

moment conditions of the form E[κiκi−j ] = 0 for j = 1, 2, 3 .As the table shows, none of the test statistics in rows (i)-(iii) are significant at the 5 percent level. In

particular, there is no evidence from the tests in row (i) that the functional forms in (7)-(9) are unduly

restrictive. The results in rows (ii) and (iii) address the question of whether there should be a distinction in

our model between the arrival of US and German news, or scheduled and unscheduled news. Recall that the

median (daily) arrival rate for German news is four times the rate for US news. Some of this difference may

be attributable to institutional features, such as the distribution of news bureaus supplying Reuters, that

are unrelated to the pace at which price-relevant information becomes known. In particular, it is possible

that the arrival rate of German news items on the Headline screens overstates the true pace at which price-

relevant German news arrives. In this case, our specification using the Ai dummy will overstate how the

dynamics of prices and order flow change immediately following the arrival of price-relevant news. The

C-statistics in row (ii) test for this form of misspecification using the arrival of US news as an instrument.

None of the statistics are significant at the 5 percent level. Differences between the arrival of scheduled and

unscheduled news could pose similar problems. For example, if the ratio of common-knowledge to dispersed

information in scheduled news is higher on average than in non-scheduled news, the price and order flow

dynamics following the arrival of scheduled news may differ from the dynamics following the arrival of other

news. The C-statistics in row (iii) are designed to look for evidence of this form of misspecification. None

are significant at the 5 percent level.14 In sum, these diagnostic tests suggest that the estimated model

adequately accounts for the effects of varying trade intensity and the arrival of news on the dynamics of

transaction prices and interdealer order flow.

The model is less successful in accounting for all the heteroskedasticity in the error processes. The C-tests

for 3rd-order residual ARCH are significant at the 5 percent level. An inspection of the estimated residuals

14Since the arrival of scheduled news is, by definition, exogenous to past market volatility, these results are consistent with theabsence of feedback from FX price volatility to the arrival of unscheduled news items. We also looked more directly for evidenceof feedback by estimating logit and probit models for Ai and Ausi and Agmi using lagged square price changes, specifically∆paski−j24j=6, as explanatory variables. In all cases, the estimated coefficients were small and statistically insignificant. Thereis no evidence of feedback effects in our filtered series of unscheduled news items.

18

shows that these residual ARCH effects are concentrated at lag one. In fact, if we omit this moment from

our C-test, we cannot reject the null of no residual heteroskedasticity. We have accounted for this feature

of the data in our estimates and tests by constructing the GMM weighting matrix from the Newey West

estimator with an MA(1) serial correlation correction.15

3.1.4 News Arrival and Intraday Dynamics

We now examine how the information in macro news is transmitted to prices. For this, we use our model

estimates to compute a variance decomposition for price changes across different market states. First, we

use our estimates to write the change in average transaction price as:

∆pi = B(L, Si)ξi + εi, (14)

where B(L, S) = D(L, S)C(L)Lm. The state-dependent coefficients in B(L, S) identify how dispersed infor-

mation affects prices and can be computed from our estimates of the coefficients inD(L, S) and C(L).We can

also use equation (14) to decompose the variance of price changes into different theoretical components. In

particular, consider the k-period price change between period i−k and i: ∆kpi ≡P k−1

j=0∆pi−j . Substituting

for ∆pi with (14), gives:

∆kpi =Xk−1

j=0εi−j +

Xk−1

j=0B(L, Si−j)ξi−j . (15)

Since the ξi, and εi shocks are mutually independent and serially uncorrelated, (15) implies that:

V ar³∆kpi| Si−jk−1j=0

´=Xk−1

j=0Ωε(Si−j) +

Xk−1

j=0B(L,Si−j)

2Ωξ(Si−j). (16)

Equation (16) provides a decomposition of the variance of price changes conditioned on the state of the

market during the last k periods. The first component on the right-hand side is the variance contribution

of common-knowledge shocks, the second is the contribution of dispersed information shocks operating via

order flow. Notice that state-dependency in the error variances and lag polynomial D(L, S) of our model

allows the contribution of each variance component to vary with changes in trade intensity and the arrival

of macro news. We now use the model estimates to quantify these effects.

Order flow is much more important in price determination when macro news arrives. Table 5 reports the

estimated contribution of dispersed information to the variance of price changes over horizons of 5, 30 and 60

minutes (i.e., k = 1, 6, 12) when trading intensity is at four different levels (i.e., n = 5, 10, 20, 30). Row (i)in each panel reports the contribution for a given level of trade intensity in the absence of macro news. (The

statistics in parenthesis are standard errors associated with these estimates computed from the asymptotic

distribution of the GMM estimates by the “delta-method”.16) Consistent with the results in Evans (2002),

15Specifically, the presence of first-order ARCH induces serial correlation in the residuals associated with conditions (10b),(11b), (11c) and (11d).16 Specifically, let Rk(θ, n,A) denote the contribution of dispersed information shocks equal to

19

these statistics show that the contribution of dispersed information to price variance rises with trade intensity

and horizon. The contribution of dispersed information in the presence of macro news is reported in row (ii).

These statistics incorporate direct effects of news arrival via the A and A dummies and the indirect effects

via the induced change in trade intensity. We estimate that trading intensity rises by approximately 9 trades

per minute when news arrives.17 To estimate the contribution of dispersed information we therefore use the

GMM estimates of (16) with B(L, Sa), Ωξ(Sa), and Ωε(Sa) where Sa = n+ 9, 1 and n is the initial level

of trade intensity shown at the top of each panel in the table. A comparison of the statistics in rows (i) and

(ii) show that following the arrival of macro news, dispersed information contributes more to the variance of

prices across all three horizons. This pattern also appears consistently across all four panels (corresponding

to different initial levels of trade intensity).

We conducted a Monte Carlo experiment to assess the statistical significance of these findings. The

experiment comprised the following steps: (i) draw a vector of parameter estimates θjfrom the estimated

asymptotic distribution of the GMM estimates; N(θ, Vθ), (ii) use (16) and θjto compute the contribution of

the dispersed information shocks to the k-period price variance at trade intensity n in the absence of news

(A = A = 0), Rk(θj, n, 0) for horizons of 5, 30 and 60 minutes (i.e., k = 1, 6, 12), (iii) use (16) and θ

jto

compute the contribution to k-period price variance with news (A = A = 1) at trade intensity na = n+ 9,

Rk(θj, na, 1) for k = 1, 6, 12 , and (iv) repeat steps (i) - (iii) 5000 times for n = 5, 10, 20, 30 and compute

the fraction of times that Rk(θj, n, 0) ≥ Rk(θ

j, na, 1). This procedure gives us a Monte Carlo estimate of

the p-value for the null hypothesis that news arrival does not increase the contribution of dispersed news to

the variance of prices. Cases where the p-values are less that 10, 5 and 1 percent are indicated in Table 5

by “∗”, “∗∗”, and “∗∗∗” respectively . Based on these calculations, the increased contribution of dispersed

information shocks following the arrival of macro news is strongly significant over most horizons and initial

trading intensities.

The specification tests reported in Table 4 do not suggest that the direct affects of macro news arrival

vary according to whether or not the news item is a scheduled. Nevertheless, scheduled news may have a

different total impact because the induced trade intensity differs from the trade intensity induced by non-

scheduled news. We estimate that trading intensity when scheduled U.S. news arrives rises by approximately

13 trades per minute. Row (iii) of Table 5 shows the contribution of dispersed information in the presence

of a scheduled news announcement that increases trade intensity by this amount. Because the price-impact

of order flow increase with trading intensity, the estimated variance contribution of dispersed information

is larger following the arrival of scheduled news than it is for the more prevalent non-scheduled items. The

nPk−1j=0 B(L,S)

2Ωξ(S)o nPk−1

j=0 Ωε(S) +Pk−1

j=0 B(L, S)2Ωξ(S)

o−1given a constant level of trading intensity n, and the

presence or absence of macro news, A = 1, 0 . We estimate the standard error of Rk(θ, n,A) as the square root of∇Rk(θ, n,A)0V∇Rk(θ, n,A) where ∇Rk(.) is the gradient vector w.r.t. θ, and V is the estimated covariance matrix of theGMM estimates, θ.17This estimate is obtained from the OLS estimate of δ from the regression: ni = δAi +

Pγidumi,τ + ui where dumi,τ is a

“seasonal” time dummy that takes the value of one when observation i falls in the τ 0th 30-minute window of a day. We estimateδ to be 8.91 with a standard error of 0.62.

20

Table 5: Variance Decomposition

Horizon (minutes) Horizon (minutes)5 30 60 5 30 60

trade intensity: n = 5 trade intensity: n = 10(i) No News 0.631 0.989 0.758 1.436 2.314 2.118

(1.040) (2.811) (3.754) (1.327) (2.911) (3.621)(ii) News 3.895∗∗ 10.280∗∗ 11.768∗ 5.123∗∗ 12.137∗∗ 13.597∗∗

(0.911) (3.396) (4.236) (1.354) (4.554) (5.451)(iii) Scheduled News 8.271∗∗∗ 16.083∗∗ 17.417∗ 9.868∗∗∗ 17.807∗∗ 19.067∗∗

(2.896) (8.020) (9.112) (3.748) (9.569) (10.727)

trade intensity: n = 20 trade intensity: n = 30(i) No News 3.808 7.475 7.957 7.173 14.862 16.129

(1.359) (2.850) (3.303) (1.738) (3.747) (4.131)(ii) News 7.981∗∗ 15.754∗ 17.101∗ 11.214∗∗ 19.163 20.358

(2.755) (7.658) (8.729) (4.500) (10.673) (11.862)(iii) Scheduled News 13.231∗∗∗ 21.067∗ 22.163∗ 16.679∗∗ 24.053 24.980

(5.533) (12.326) (13.573) (7.248) (14.569) (15.871)

Notes: The table reports values for Rk(θ, n,A), the contribution of dispersed information shocksto variance of k-horizon price changes implied by the GMM estimates of the intraday model given aconstant level of trading intensity n, and the presence or absence of macro news, A = 1, 0 .Standard errors are in parentheses. Statistics in rows (i) - (iii) are computed as Rk(θ, n, 0),Rk(θ, n + 9, 1) and Rk(θ, n + 13, 1) respectively. Cases where the Monte Carlo p-value for thenull that news arrival does not increase the contribution of dispersed news to the variance of pricesis less than 10, 5 and 1 percent are indicated by “∗”, “∗∗”, and “∗∗∗” respectively.

p-values computed from Monte Carlo experiments with na = n + 13 indicate an even stronger pattern of

statistical significance.

Overall, our estimates indicate that order flow contributes more to price adjustment following macro news

than at other times. This is not what one would expect if macro news is primarily comprised of common-

kowledge information that is directly impounded into FX prices. If macro news primarily transmits new

common-knowledge information, order flow should contribute less to price-dynamics in the period following

the arrival of news than at other times. By contrast, the results in Table 5 strongly suggest that the arrival

of macro news triggers trading that reveals new dispersed information that affects prices indirectly. One

particularly interesting aspect of our findings concerns the effects of scheduled U.S. announcements. Since

these news items contain data releases on macro economic aggregates, one might have expected that they

contain a greater proportion of common-knowledge to dispersed information than some of the other news

items in our sample. That order flow is at least as important in price dynamics following scheduled news

suggests that this common view concerning the information content of macro news is incorrect.

21

4 Daily Analysis

Our intraday analysis shows the importance of the order flow channel as a means for impounding macro

news in FX prices. We now examine implications of this for the behavior of FX prices at the daily frequency.

This examination compliments our intraday analysis for three reasons. First, daily changes in FX prices are

very nearly a martingale (which is not true of five-minute changes). Our daily model thus sheds light on

how the information contained in macro news contributes to price variation over the longer run. Second,

our daily analysis provides additional perspective on results relating daily price dynamics to order flow (e.g.,

Evans and Lyons 2002a). In particular, our estimates provide a breakdown of the sources of price and order

flow volatility. Third, our daily analysis provides a robustness check on the results presented above. For

example, we can construct measures of the daily flow of macro news in ways that were not possible at higher

frequencies. The consistency of the results derived from estimates of the daily and intraday model shows

that our main findings are robust to our methods for identifying the impact of macro news arrivals.

4.1 The Model

Our daily model for price and order flow dynamics comprises the following equations:

∆pt = αxt + et + vt, (17)

xt = ut + wt, (18)

where ∆pt is the change in the spot price of FX between 5:00 pm on day t− 1 and 5:00 pm on day t and xt

is interdealer order flow realized over the same period. The parameter α captures the price impact of order

flow at the daily horizon, i.e., it reflects information content. Prices and order flow are subject to four shocks

representing different sources of information hitting the market: et, vt, ut, and wt. These shocks are mean

zero, mutually uncorrelated, and serially uncorrelated. The et and vt shocks represent information that is

impounded in price directly. et is the common knowledge effect of macro news arrivals on the price of FX.

vt represents other factors directly impounded in prices, i.e., factors unrelated to both order flow or macro

news events (possibly noise). Order flow is driven by the ut and wt shocks. The ut shocks represent order

flow effects from macro news arrivals — the dispersed information effect of the news. Shocks to order flow

that are unrelated to macro news are represented by the wt shocks (e.g., portfolio shifts arising from other

sources such as changing risk tolerances or hedging).

We identify the effects of the news-related common-knowledge and dispersed-information shocks, et and

ut, through state-dependency of price changes and order flow in the second moments. Specifically, we assume

that the variance of et and ut on day t is increasing in the daily flow of macro news, which we measure by

the number of US and German news arrivals between 5:00 pm on days t− 1 and t, Aust and Agt :

V art (et) = Σ2e(A

ust , Agt ), and V art (ut) = Σ

2u(A

ust , Agt ), (19)

22

where Σ2κ (0, 0) = 0, with ∂Σ2κ /∂Akt > 0 for κ = e, u and k = us,g . Thus, on days without news,

et = ut = 0, so price changes and order flow are driven solely by the vt and wt shocks. These shocks are

independent of news, so their variances are unrelated to Akt . As we shall see, there is little evidence of

state-dependency in the second moments of daily price changes and order flow beyond the effects of news. In

particular, unlike our intraday model, there is no need to incorporate trade intensity as an additional state

variable. We therefore assume that the conditional variances of the vt and wt shocks are constant:

V art (vt) = Σ2v, and V art (wt) = Σ

2w. (20)

Several features of our daily model deserve comment. First, our specification abstracts from the complex

intraday dynamics of prices and order flow. Equations in (17) and (18) imply that by 5:00 pm GMT each

day, FX prices fully reflect the information contained in order flow to that point. As a result, price change

over the next 24 hours (i.e. ∆pt+1) are not correlated with order flow from the past 24 hours, (i.e. xt).

This feature of our model is supported by the data. We show below that there is no correlation between

∆pt+1 and xt. Our specification also implies the absence of serial correlation in daily price changes and order

flows. This too is consistent with the evidence reported in Section 2. A second feature of our specification

concerns the price-impact parameter α. Our intraday analysis showed that the price impact of order flows

varied with trade intensity and the arrival of news. This form of state-dependency in the intraday data does

not appear at the daily frequency (addressed below), so we do not allow for state-dependency in α.We would

add that this restriction in our model means that our test of the relative importance of indirect effects is

conservative: order flow induced by news may have more price impact than the constrained equation gives

it credit for. In any event, we do incorporate state-dependency into the error variances. This final feature is

key to identifying the effects of macro news, so let us focus on it more closely.

Identification of the effects of macro news is achieved by the assumption that the variance of the et and

ut shocks is higher on days when there are a greater number of news items appearing on the Reuters Money

Market News screen. Crucially, this assumption does not require that FX market participants view the

information in each news item as equally important (which the market does not). The identifying power of

this assumption does, however, depend on the absence of wild variations in the quality of Reuters’ editorial

judgements. For example, if the Reuters screen were flooded one day with reports containing essentially

no information, but on another a few reports appeared with great economic significance, daily variations

in the number of news reports would be a poor measure of the daily flow of macro news. Based on our

understanding of Reuters’ editorial process, this possibility seems far-fetched. That said, we recognize that

no single measure will identify the daily variation in macro news flow with complete precision. Thus, in

addition to measures based on the daily arrival rates for US and German news shown in (19), we will also

use measures based on the subset of items that are scheduled.

23

4.2 Estimation

We estimate two versions of the model by the Generalized Method of Moments. Version I assumes that

the variance of the et and ut shocks on day t varies only with the sum of the US and German news items,

Aallt ≡ Aust +Agt . Under this specification, the flow of macro news is identified by the arrival rate of both US

and German news. We also allow for the possibility that daily variations in the flow of macro news may be

reflected differently in the arrival rates for US and German news. Version II of our model allows the variance

of et and ut on day t to depend on the number of US and German news items separately. The variance

functions are assumed to be linear in both versions of the model:

Version I: Σ2κ (Aust , Agt ) = σκA

allt

Version II: Σ2κ (Aust , Agt ) = σusκ A

ust + σgκA

gt

(21)

where σκ , σusκ and σgκ are positive parameters for κ = e, u . Thus, the parameters to be estimated areα,Σ2w,Σ2v, σe, σu in Version I, and α,Σ2w,Σ2v, σuse , σge , σ

usu , σgu in Version II.

The GMM estimates of the model parameter are derived from the following set of moment conditions:

0 = E [(∆pt − αxt)xt] (22a)

0 = E [Vt (∆pt)− V art (∆pt)⊗Zt] , (22b)

0 = E [Vt (xt)− V art (xt)⊗Zt] , (22c)

where Zt is a vector of instruments. Condition (22a) follows from the assumed orthogonality between the

shocks to prices (et and vt) and the shocks to order flow (ut and wt). Conditions (22b) and (22c) combine

the second moments of price changes and order flow implied by the model with measures of the variance of

order flow, V (xt), the variance of price changes, V (∆pt) . These measures are computed for each day in oursample from the 5-minute intraday observations as:

Vt (∆pt) =288Xi=1

∆p2it, Vt (xt) =288Xi=1

x2it, (23)

where the subscript “it” denotes the i0th 5-minute observation on day t. Vt (∆pt) and Vt (xt) are the (un-centered) second moments of the price change and order flow process over day t, scaled by the number of

5-minute intraday observations. Andersen, Bollerslev, Diebold, and Labys (2001) show that these measures

are consistent nonparametric estimates of the actual moments under mild regularity conditions. They also

note that while the measures will be biased when prices changes and order flow do not follow Martingales

in the continuous time limit, in practice these biases will be very small if a large number of high frequency

observations are used to compute each daily measure. This appears true in our data. Estimates of Vt (∆pt) ,and Vt (xt) computed from ∆pit and xit are almost identical to their counterparts using the estimated resid-

24

uals from the price and order flow equations of the intraday model: the correlation between the alternative

measures is greater than 0.99 for both order flow and price changes.18

We use two sets of instruments to implement estimation. The instrument vector in Version I comprises

a constant and sum of the US and German news items, Aallt . In Version II, we use a constant, Aust and Agt

as instruments. These choices imply that the number of moment conditions in (22a-c) equals the number

of parameters, so the estimates come from exactly identified versions of model. As above, we apply the

standard 2-step method to compute the GMM estimates (without the serial correlation correction in the

weighting matrix). We will also consider the adequacy of our model estimates with a set of diagnostic tests

based on additional moment conditions.

4.2.1 Daily Estimates

Panel A of Table 6 reports parameter estimates from both versions of the model with exact identification.

Asymptotic standard errors allowing for residual heteroskedasticity are shown in parentheses. In both spec-

ifications the estimate of the price-impact parameter α is positive, as the theory predicts, and statistically

significant. (Its size corresponds to a price impact of roughly 50 basis points per $1 billion in order flow.) In

Version I of the model, both variance parameters σe and σu are positive and significant at the five percent

level. These estimates imply that both direct and indirect effects of news on price are present. This finding

is confirmed by the estimates from Version II reported in the right-hand panel. When U.S. and German

news events are introduced separately, the estimates of σuse , σge , σusu , and σgu are all positive and significant

at the five percent level. Furthermore, as panel B shows, Wald statistics for the null that σuse = σusu = 0,

and σusu = σgu = 0, are highly significant. Panel B also shows that there is no significant evidence against the

parameter restrictions imposed by Version I of the model, namely σuse = σusu and σusu = σgu.

Panel C shows results of diagnostic tests that examine an expanded set of moment conditions provide

additional support for our specification. In row (i) we report the J-statistic for specifications using (22)

and E [(∆pt − αxt)xt−1] = 0 as moment conditions.19 Our model should satisfy this additional condition

because all the price impact of order flow occurs within the day. As the table shows, there is no signif-

icant evidence to reject this set of restrictions in either version of the model. The statistics in row (ii)

test for the presence of (residual) serial correlation in the price change and order flow process by respec-

tively adding E [(∆pt − αxt) (∆pt−1 − αxt−1)] = 0 and E [xtxt−1] = 0 to the conditions in (22). Again,

consistent with the assumed structure of out model, none of the J-statistics are statistically significant.

Next, we turn to the issue of state-dependency. Our daily model assumes that trade intensity and news

have no effect on α, the parameter identifying the price-impact of order flow. We examine this restric-

tion by adding E [(∆pt − αxt)⊗ zt] = 0 to the conditions in (22) for zt = xtnt, xtAallt in Version I and

18This finding is not inconsistent with the results of intraday model. Although intraday order flow is serially correlated,lagged order flow accounts for a very small faction of the variation in realized order flow over a 5-minute period. Similarly, theleads and lags in intraday order flow account for a small fraction of the variance in high frequency price changes.19The J-statistics reported here are equivalent to the C-statistics used in our intraday analysis because both versions of the

daily model are exactly identified without the additional moment conditions.

25

Table 6: GMM Estimates of Daily Models

A: Parameters Version I Version IIEstimate Std. Err Estimate Std. Err

α 0.032 (0.003) 0.032 (0.003)σ2w 67.231 (11.395) 67.018 (11.282)σ2v 3.530 (0.675) 3.518 (0.671)σe 3.737 (0.813)σu 0.188 (0.053)σuse 5.682 (2.661)σge 3.358 (0.977)σusu 0.291 (0.147)σgu 0.168 (0.063)

B: Wald Tests Statistic p-value

σuse = σusu = 0 33.303 (0.000)σusu = σgu = 0 13.707 (0.001)σuse = σusu & σusu = σgu 0.763 (0.683)

C: Diagnostic Tests Statistic p-value Statistic p-value

i) Lagged order flow 2.502 (0.114) 2.502 (0.114)ii) Serial correlation:

∆pt eqn. 0.014 (0.905) 0.014 (0.905)xt eqn. 0.190 (0.663) 0.190 (0.663)

iii) State-dependency:α 2.767 (0.251) 2.767 (0.251)Var(∆pt) & Var(xt) 2.479 (0.290) 2.527 (0.283)

iv) Residual Arch:∆pt eqn. 0.348 (0.555) 0.281 (0.596)xt eqn. 2.332 (0.127) 2.486 (0.115)

v) Joint Test 10.097 (0.343) 9.876 (0.361)Notes: Panel A of the table reports GMM parameter estimates and asymptotic standarderrors (corrected for heteroskedasticity) in parentheses. Panel B shows Wald tests forthe coefficient restrictions listed on the left with asymptotic p-values reported in paren-theses. The J−tests shown in panel C test the moment restrictions in (22) and thefollowing: (i) E [(∆pt − αxt)xt−1] = 0, (ii) E [(∆pt − αxt) (∆pt−1 − αxt−1)] = 0,E [xtxt−1] = 0, (iii) E [(∆pt − αxt)⊗ zt] = 0, E [Vt (∆pt)− V art (∆pt)nt] = 0,and E [Vt (xt)− V art (xt)nt] = 0, where zt = xtnt, xtAallt in Version I and zt =xtnt, xtAust , xtA

gt in Version II, (iv) E[Vt (∆pt)− V art (∆pt) Vt−1 (∆pt−1) −

V art−1 (∆pt−1)] = 0 and E[Vt (xt)− V art (xt) Vt−1 (xt−1)− V art−1 (xt−1)] =0, and (v) all the moments listed in (i) - (iv). Asymptotic p-values are reported in paren-theses.

26

zt = xtnt, xtAust , xtAgt in Version II, where nt denotes trading intensity on day t. As the table shows, nei-

ther of the associated J−statistics are significant. We also check for additional state-dependency in the errorvariances. In this case we add E [Vt (∆pt)− V art (∆pt)nt] = 0 and E [Vt (xt)− V art (xt)nt] = 0 to theconditions in (22). These additional moments examine whether the residual variance in price and order flow,

unaccounted for by the arrival of news, is correlated with daily trade intensity. Once again, neither of the

J-statistics is significant. There is no evidence that trade intensity should be present as a second state vari-

able governing the error variances. Further evidence on the specification of the error variances is provided by

the statistics in row (iv). Here we test for residual first order ARCH by adding E[Vt (∆pt)− V art (∆pt)Vt−1 (∆pt−1) − V art−1 (∆pt−1)] = 0 and E[Vt (xt)− V art (xt) Vt−1 (xt−1)− V art−1 (xt−1)] = 0 to

the conditions in (22). These specification tests also show no evidence of significant misspecification in the

error variances.20 Finally, in row (v), we report J−statistics for models using (22) and all the additionalmoments. These moment conditions respectively provide 9 and 11 over-identifying restrictions in Versions I

and II of the model. As the table shows, neither J−statistic is significant at the 5 percent level. The para-meter estimates obtained in this manner are very similar to those reported in Panel A. Since the estimated

standard errors are a little smaller (as one would expect), the overall pattern of statistical significance we

report appears robust to the number of over-identifying restrictions used in estimation. Importantly this

level of robustness is also reflected in the model-based statistics we consider next.

4.3 News Arrival and Daily Dynamics

Our intraday analysis showed that dispersed information contributes more to the variance of price changes

following macro news announcements than at other times. Our daily model allows us to address a distinct

but equally important issue: the extent to which macro news is impounded in prices directly, via the common

knowledge et shocks, or indirectly via the dispersed information ut shocks that affect prices via order flow.

To clarify this issue within the context of our daily model, consider the unconditional variance of price

changes implied by our model, V ar (∆pt). By definition, this variance can be written as E [V art (∆pt)] +

V ar (Et∆pt) where Et∆pt and V art(∆pt) denote the first and second moments of price changes conditioned

on the day t state of the market. According to our model, the number of news arrivals has no implication

for the direction of how prices will change, so Et∆pt = 0. With the aid of equation (17), we can therefore

write the unconditional variance as:

V ar (∆pt) = α2E [V art (xt)] +E [V art (et + vt)] .

20An earlier version of this paper examined two further aspects of the model. We looked for evidence of nonlinearity inthe error-variance specifications shown in (21) by regressing Vt (∆pt) and Vt (xt) on a constant, Aust , Agt , (A

ust )

2, and (Agt )2.

Since the price and order flow variances are linear functions of the error variances, nonlinearity in the latter should appearin the form of non-zero coefficients on (Aust )

2 and (Agt )2 in these regressions. Our estimates of these coefficients were not

statistically significant. We also explored whether temporal aggregation could affect our results by introducing a feedback fromprice changes to order flow. Model estimates incorporating this feedback effect were similar to those reported here, and hadthe same implications concerning the effects of macro news.

27

The first term on the right identifies the contribution of order flow volatility to the variance of price changes.

The second term identifies the contribution of information that is directly impounded into prices. Using

equations (18)-(20) to substitute for V art (xt) and V art (et + vt) , we obtain:

V ar (∆pt) = E£Σ2e(A

ust , Agt )

¤+ α2E

£Σ2u(A

ust , Agt )

¤+Σ2v + α2Σ2w. (24)

Equation (24) decomposes the unconditional variance of daily price changes into four components. The

first term identifies the contribution of common-knowledge shocks associated with the arrival of news. We

refer to this as the direct channel. The second term represents the contribution of dispersed information

shocks associated with news. Notice that this term includes the price-impact coefficient α, because dispersed

information affects prices via order flow. We refer to this as the indirect channel. The third and fourth terms

identify the contribution of shocks that are not associated with the arrival of news; information embedded

in the vt and wt shocks affects price via the direct and indirect channels respectively.

Table 7 reports elements of the variance decomposition in (24) derived from the estimates of the daily

model. For this purpose, the expectations terms in (24) are replaced by sample averages (i.e., E£Σ2κ (A

ust , Agt )

¤is replaced by 1

T

PTt=1 Σ

2κ (A

ust , Agt ) for κ = e, u). We also report standard errors computed by the “delta-

method” from the estimated asymptotic distribution of the model estimates. The statistics shown in Panel

A use the parameters estimated from the exactly identified models reported in panel A of Table 6. As noted

above, these statistics are very similar to those based on the estimates derived from Versions I and II of the

model with 9 and 11 over-identifying restrictions.

The upper rows in panel A of Table 7 report the contribution of dispersed and common knowledge

information shocks to the unconditional variance of prices. The statistics in row (i) report the frac-

tion of the unconditional variance attributable to the common knowledge shocks associated with news:

E[Σ2e(Aust , Agt )]/V ar (∆pt). Estimates from both versions of the model indicate that the direct effect of news

arrivals account for approximately 14 percent of the variance of total price changes. The estimates from

Version II of the model indicate that this total is split roughly 2 to 1 between German and US news. Since

German news arrives at four times the daily rate of US news on average, these estimates suggest that a

typical US news item has a somewhat larger direct effect on prices than a German item. Row (ii) reports

the contribution of dispersed information to the variance of prices: α2E[Σ2u(Aust , Agt )]/V ar (∆pt). These

statistics show that the indirect effects of news arrival account for roughly 22 percent of the variance. Once

again, the arrival of German news contributes more than twice as much as US news through this channel.

Row (iii) shows the total contribution of news to the variance of prices via both channels is approximately

36 percent. These estimates are an order of magnitude larger than those found in event studies. As noted

in the introduction, the results from these studies imply that scheduled announcements account for only 2

to 3 per cent of the variance in daily price changes. The two main differences in approach are that here we

consider both direct and indirect channels and we use a larger set of news items. Row (iv) shows the relative

importance of the latter. Here we report the ratio of indirect to direct effects of news arrival implied by our

28

Table 7: Daily Price Variance Decompositions

A: All News Version I Version II

Combined US German Combined

(i) Direct 0.139 0.036 0.104 0.140(0.046) (0.042) (0.017) (0.046)

(ii) Indirect 0.224 0.060 0.166 0.226(0.078) (0.033) (0.070) (0.078)

(iii) Total 0.364 0.096 0.270 0.366(0.092) (0.040) (0.088) (0.091)

(iv) Ratio (Indirect/Direct) 1.612 1.642 1.602 1.612(0.763) (1.069) (0.857) (0.761)

B: Scheduled News I Version I Version II

Combined US German Combined

(v) Total 0.063 0.067 0.027 0.093(0.036) (0.037) (0.023) (0.046)

(vi) Ratio (Indirect/Direct) 1.399 1.223 1.589 1.317(1.060) (0.893) (1.945) (0.855)

C: Scheduled News II Version I Version II

Combined US German Combined(vii) Total 0.038 0.018 0.024 0.042

(0.034) (0.017) (0.028) (0.037)(viii) Ratio (Indirect/Direct) 1.249 1.646 1.105 1.303

(1.040) (2.486) (1.288) (1.093)Notes: The table reports elements of the variance decomposition for price changesimplied by the GMM estimates of the daily models. Rows (i) - (iv) re-port estimates of E[Σ2e(At)]/V ar (∆pt) , α

2E[Σ2u(At)]/V ar (∆pt) , (E[Σ2e(At)] +α2E[Σ2u(At)])/V ar (∆pt) and α2E[Σ2u(At)]/E[Σ

2e(At)]. Under the Version I heading,

the estimates use At = Aallt . Under the US, German and Combined headings of VersionII, At equals Aust , Agt and A

allt . Panels B and C respectively report estimates using the

number of scheduled news items, and the absolute forecast error for scheduled news. Stan-dard errors computed from the estimated asymptotic distribution of the GMM estimatesare reported in parentheses.

model estimates: α2E[Σ2u(Aust , Agt )]/E[Σ

2e(A

ust , Agt )]. As the table shows, the contribution of news via the

indirect channel is roughly 60 percent larger than the contribution via the direct channel. These estimates

clearly indicate that the indirect effects of news operating via order flow are an important component of

price dynamics.

Panels B and C of Table 7 are valuable for evaluating the robustness of our finding that indirect effects

29

are relatively important. The panels verify that the same conclusion arises across quite different definitions

of news. In panel B we report elements of the unconditional price variance derived from Versions I and II of

our daily model estimated with scheduled news only. This set includes US scheduled news concerning Non

Farm Payroll, Durable Goods, Trade Balance, and Unemployment Claims, and all the German news items

covered by MMS (see Section 2 above). As the table shows, estimates from Versions I and II of this modified

model imply that scheduled news contributes between 6 and 10 per cent to the unconditional variance of

prices. While these estimates are only about one quarter the size of their counterparts based on the full

spectrum of news in panel A, they are still somewhat larger than the contribution implied by event studies.

The ratio statistics reported in row (vi) suggest why this might be so. Here we see that, in sharp contrast to

textbook models, scheduled news affects price more through the indirect channel than directly. It appears

that scheduled news triggers trading that reveals incremental dispersed information.

As a further robustness check, we re-estimated the model using same set of scheduled news in a different

way. The form of the model remains unchanged except that Aust and Agt in (21) now denote the sum of

absolute standardized forecast errors for scheduled US and German news on day t. The forecast errors are

computed from the surveys conducted by MMS, and the errors for each series are standardized using the

standard deviation of the errors from January 1993 to December 1999.21 Panel C of Table 7 reports elements

of the price variance derived from estimates of the daily model using this measure of scheduled news arrival.

Here we see that scheduled news contributes about 4 percent to the variance of prices, while the ratios of

indirect to direct effects remain above one.

To summarize, the results in Table 7 show that both scheduled and non-scheduled news contribute to the

variance of the price changes in our sample. Our results also indicate that news items generally contain both

common-knowledge information that is directly reflected in prices, and dispersed information that indirectly

affects prices via its impact on order flow. Furthermore, all our model estimates imply that this indirect

channel is at least as important as the direct channel in linking news to price changes.

5 Conclusion

This paper extends past work on FX prices and public news in three main ways. We address the presence

of an indirect channel through which public news affects prices. Second, we use heteroskedasticity in order

flow and price for identification, à la Rigobon and Sack (2004), rather than the more common event-study

approach. Third, our methodology exploits the full set of macro news events piped into FX trading desks.

Our analysis of intraday data shows that order flow contributes more to changing FX prices in the period

immediately following the arrival of news than at other times. This evidence pointing to the importance of

the indirect channel is supported by our daily analysis: roughly two-thirds of the effect of macro news on FX

21Love and Payne (2004) construct a similar aggregate measure except that they “sign” each forecast error according to thedirection of its theoretically predicted exchange rate effect. The latter adjustment is unnecessary here because our aim is toidentify changes in the flow of macro news rather than to identify the directional influence of scheduled news on FX prices.

30

prices is transmitted via order flow, the remainder being the direct effect of news. With both the direct and

indirect channels operating, we estimate that macro news accounts for 36 percent of total FX price variance

in daily data. Given that daily prices are very nearly a martingale, this finding implies that macro news is

far larger contributor to longer term price variation than previously thought.

Our daily results speak directly to the question, What drives order flow? The analysis in Evans and

Lyons (2002a) splits total daily DM/$ price variation into two parts: about 60 percent is due to order flow

and about 40 percent is due to other factors. The results in Table 7 shed light on both of these parts. They

suggest that order flow’s 60 percent breaks roughly into one-third (20 percent) that is induced by macro

news and two-thirds (40 percent) that is not news induced. Put differently, macro news accounts for about

one-third of the variance of interdealer order flow in our sample. The 40 percent of total price variation due

to other factors breaks into about one-third 15 percent from the direct effect of macro news and two-thirds

(25 percent) that remains unaccounted for.

Finally, let us offer a wider perspective on our results. Inherent in current macro models is the view

that price-setting marketmakers observe macro news, calculate the price implication, and instantly adjust

all their FX prices by the same amount. Our results suggest that this is over-simplified. Rather, they

suggest a model in which marketmakers observe macro news but have little idea how to interpret it, or

how the rest of the market will interpret it. Instead, they wait to observe the trades induced and set their

prices and expectations based on the interpretations embedded therein. (This view is consistent with the

findings of Evans and Lyons 2005 that FX order flow conveys incremental information useful for forecasting

macro variables.) Models with this richer informational structure may offer new insights into many of the

long-standing puzzles concerning the behavior of FX prices.

31

References

Andersen, T., and T. Bollerslev (1998), Deutsche mark-dollar volatility: Intraday activity patterns, macro-

economic announcements, and longer run dependencies, Journal of Finance, 53: 219-265.

Andersen, T., T. Bollerslev, F. Diebold, C. Vega (2003), Micro effects of macro announcements: Real-time

price discovery in foreign exchange, American Economic Review, 93: 38-62.

Andersen, T., T. Bollerslev, F. Diebold, and P. Labys (2001), The distribution of realized exchange rate

volatility, Journal of the American Statistical Association, 96: 42-55.

Balduzzi, P., E. Elton, and C. Green (2001), Economic news and bond prices: Evidence from the U.S.

Treasury Market, Journal of Financial and Quantitative Analysis, 36: 523-543.

Berry, T., and K. Howe (1994), Public information arrival, Journal of Finance, 49: 1331-1346.

Bollerslev, T., J. Cai, and F. Song (2000), Intraday periodicity, long-memory volatility, and macroeconomic

announcement effects in the U.S. treasury bond market, Journal of Empirical Finance, 7: 37-55.

Brandt, M., and K. Kavajecz (2004), Price discovery in the U.S. treasury market: The impact of order flow

and liquidity on the yield curve, Journal of Finance, 59: 2623-2654.

Cao, H., M. Evans, and R. Lyons (2006), Inventory information, Journal of Business, 79: 325-364.

Carlson, J., and M. Lo (2004), One minute in the life of the DM/$: Public news in an electronic market,

Journal of International Money and Finance, forthcoming.

Cornell, B. (1982), Money supply announcements, interest rates, and foreign exchange, Journal of Interna-

tional Money and Finance, 1: 201-208.

DeGennaro, R., and R. Shrieves (1997), Public information releases, private information arrival, and volatil-

ity in the foreign exchange market, Journal of Empirical Finance, 4: 295-315.

Eddelbuttel, D., and T. McCurdy (1998), The impact of news on foreign exchange rates: Evidence from

high frequency data, typescript, University of Toronto.

Ederington, L., and J. Lee (1995), The short-run dynamics of price adjustment to new information, Journal

of Financial and Quantitative Analysis, 30: 117-134.

Engel, C., and J. Frankel (1984), Why interest rates react to money announcements: An answer from the

foreign exchange market, Journal of Monetary Economics, 13: 31-39.

Evans, M. (2002), FX trading and exchange rate dynamics, Journal of Finance, 57: 2405-2448.

Evans, M., and R. Lyons (2002a), Order flow and exchange rate dynamics, Journal of Political Economy,

110: 170-180.

32

Evans, M., and R. Lyons (2002b), Time-varying liquidity in foreign exchange, Journal of Monetary Eco-

nomics, 49: 1025-1051.

Evans, M., and R. Lyons (2005), Exchange rate fundamentals and order flow, typescript, U.C. Berkeley,

September.

Fleming, M., (2003), Measuring treasury market liquidity, Federal Reserve Bank of New York Economic

Policy Review, 9: 83-108.

Fleming, M., and E. Remolona (1997), What moves the bond market? Federal Reserve Bank of New York

Economic Policy Review, 3: 31-50.

Fleming, M., and E. Remolona (1999), Price formation and liquidity in the U.S. treasury market, Journal

of Finance, 54: 1901-1915.

French, K., and R. Roll (1986), Stock return variance: The arrival of information and the reaction of

traders, Journal of Financial Economics, 17: 5-26.

Froot, K., and T. Ramadorai (2005), Currency returns, intrinsic value, and institutional-investor flows,

Journal of Finance, 60: 1535-1566.

Glosten, L., and P. Milgrom (1985), Bid, ask, and transaction prices in a specialist market with heteroge-

neously informed agents, Journal of Financial Economics, 14: 71-100.

Goodhart, C., S. Hall, S. Henry, and B. Pesaran (1993), News effects in a high-frequency model of the

sterling-dollar exchange rate, Journal of Applied Econometrics, 8: 1-13.

Green, C. (2004), Economic news and the impact of trading on bond prices, Journal of Finance, 59: 1201-

1234.

Hakkio C., and D. Pearce (1985), The reaction of exchange rates to economic news, Economic Inquiry, 23:

621-635.

Hansen, L. (1982), Large sample properties of generalized method of moments estimators, Econometrica,

50: 1029-1054.

Hardouvelis, G. (1988), Economic news, exchange rates, and interest rates, Journal of International Money

and Finance, 7: 23-25.

Hasbrouck, J. (1991), Measuring the information content of stock trades, Journal of Finance, 46: 179-207.

Hayashi, F. (2000) Econometrics, Princeton University Press, Princeton NJ.

Huang, R., J. Cai, and X. Wang (2002), Information-based trading in the interdealer market, Journal of

Financial Intermediation, 11: 269-296.

33

Ito, T., and V. Roley (1987), News from the U.S. and Japan: Which moves the yen/dollar exchange rate?

Journal of Monetary Economics, 19: 255-277.

Kandel, E., and N. Pearson (1995), Differential interpretation of public signals and trade in speculative

markets, Journal of Political Economy, 103: 831-872.

Kim, O., and R. Verrecchia (1994), Market liquidity and volume around earnings announcements, Journal

of Accounting and Economics, 17: 41-67.

Kim, O., and R. Verrecchia (1997), Pre-announcement and event-period information, Journal of Accounting

and Economics, 24: 395-419.

Klein, M. (1991), Managing the dollar: Has the Plaza Agreement mattered? Journal of Money, Credit,

and Banking, 23: 742-751.

Kyle, A. (1985), Continuous auctions and insider trading, Econometrica, 53: 1315-35.

Love, R., and R. Payne (2004), Macroeconomic news, order flows, and exchange rates, typescript, London

School of Economics.

Lyons, R. (2001), The Microstructure Approach to Exchange Rates, MIT Press.

Madhavan, A. (2000), Market microstructure: A survey, Journal of Financial Markets, 3: 205-258.

Madhavan, A., M. Richardson, and M. Roomans (1997), Why do security prices change? A transaction-level

analysis of NYSE stocks, Review of Financial Studies, 10: 1035-1064.

Melvin, M., and X. Yin (2000), Public information arrival, exchange rate volatility, and quote frequency,

Economic Journal, 110: 644-661.

Newey, W., and K. West (1987), A simple positive semi-definite, heteroskedasticity and autocorrelation

consistent covariance matrix, Econometrica, 55: 703-708.

Payne, R. (2003), Informed trade in spot foreign exchange markets: An empirical analysis, Journal of

International Economics, 61: 307-329.

Potter, S. (1999), Non-linear time series modelling: An introduction, typescript, Federal Reserve Bank of

New York.

Rigobon, R., and B. Sack, (2004), The impact of monetary policy on asset prices, Journal of Monetary

Economics, 51: 1553-1575.

Rime, D. (2000), Private or public information in foreign exchange markets? An empirical analysis, type-

script, University of Oslo, March.

34